Skip to contentSkip to conversation

RDMA/rxe: For invalidate compare keys according to the MR access

PATCH In review Md Haris Iqbal 3 participants 11 messages Jun 29, 2022 – Jul 7, 2022

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 RDMA/rxe: For invalidate compare keys according to the MR access

In rxe, the access permissions decide which of the lkey/rkey would be set during an MR registration. For an MR with only LOCAL access, only lkey is set and rkey stays 0. For an MR with REMOTE access, both lkey and rkey are set, such that rkey=lkey.

Hence, for MR invalidate, do the comparison for keys according to the MR access. Since the invalidate wr can contain only a single key (ex.invalidate_rkey), it should match MR->rkey if the MR being invalidated has REMOTE access. If the MR has only LOCAL access, then that key should match MR->lkey.

Fixes 3902b429ca14 ("RDMA/rxe: Implement invalidate MW operations")
---
 drivers/infiniband/sw/rxe/rxe_loc.h |  2 +-
 drivers/infiniband/sw/rxe/rxe_mr.c  | 17 +++++++++++------
 2 files changed, 12 insertions(+), 7 deletions(-)
drivers/infiniband/sw/rxe/rxe_loc.h +1 -1
index 0e022ae1b8a5..37484a559d20 100644
@@ -77,7 +77,7 @@ struct rxe_mr *lookup_mr(struct rxe_pd *pd, int access, u32 key,
 			 enum rxe_mr_lookup_type type);
 int mr_check_range(struct rxe_mr *mr, u64 iova, size_t length);
 int advance_dma_data(struct rxe_dma_info *dma, unsigned int length);
-int rxe_invalidate_mr(struct rxe_qp *qp, u32 rkey);
+int rxe_invalidate_mr(struct rxe_qp *qp, u32 key);
 int rxe_reg_fast_mr(struct rxe_qp *qp, struct rxe_send_wqe *wqe);
 int rxe_mr_set_page(struct ib_mr *ibmr, u64 addr);
 int rxe_dereg_mr(struct ib_mr *ibmr, struct ib_udata *udata);
drivers/infiniband/sw/rxe/rxe_mr.c +11 -6
index fc3942e04a1f..790cff7077fd 100644
@@ -576,22 +576,27 @@ struct rxe_mr *lookup_mr(struct rxe_pd *pd, int access, u32 key,
 	return mr;
 }
 
-int rxe_invalidate_mr(struct rxe_qp *qp, u32 rkey)
+int rxe_invalidate_mr(struct rxe_qp *qp, u32 key)
 {
 	struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
 	struct rxe_mr *mr;
 	int ret;
 
-	mr = rxe_pool_get_index(&rxe->mr_pool, rkey >> 8);
+	mr = rxe_pool_get_index(&rxe->mr_pool, key >> 8);
 	if (!mr) {
-		pr_err("%s: No MR for rkey %#x\n", __func__, rkey);
+		pr_err("%s: No MR for key %#x\n", __func__, key);
 		ret = -EINVAL;
 		goto err;
 	}
 
-	if (rkey != mr->rkey) {
-		pr_err("%s: rkey (%#x) doesn't match mr->rkey (%#x)\n",
-			__func__, rkey, mr->rkey);
+	if ((mr->access & IB_ACCESS_REMOTE) && key != mr->rkey) {
+		pr_err("%s: key (%#x) doesn't match mr->rkey (%#x)\n",
+			__func__, key, mr->rkey);
+		ret = -EINVAL;
+		goto err_drop_ref;
+	} else if ((mr->access & IB_ACCESS_LOCAL_WRITE) && key != mr->lkey) {
+		pr_err("%s: key (%#x) doesn't match mr->lkey (%#x)\n",
+			__func__, key, mr->lkey);
 		ret = -EINVAL;
 		goto err_drop_ref;
 	}
Signature · 1 line
2.25.1
#1 lore raw
Quoted 6 lines · Md Haris Iqbal

In rxe, the access permissions decide which of the lkey/rkey would be set during an MR registration. For an MR with only LOCAL access, only lkey is set and rkey stays 0. For an MR with REMOTE access, both lkey and rkey are set, such that rkey=lkey.

Quoted 7 lines

Hence, for MR invalidate, do the comparison for keys according to the MR access. Since the invalidate wr can contain only a single key (ex.invalidate_rkey), it should match MR->rkey if the MR being invalidated has REMOTE access. If the MR has only LOCAL access, then that key should match MR->lkey.

Fixes 3902b429ca14 ("RDMA/rxe: Implement invalidate MW operations")

I think I messed up the Fixes line. It should be,

Fixes 001345339f4c ("RDMA/rxe: Separate HW and SW l/rkeys")
Quoted 61 lines
---
 drivers/infiniband/sw/rxe/rxe_loc.h |  2 +-
 drivers/infiniband/sw/rxe/rxe_mr.c  | 17 +++++++++++------
 2 files changed, 12 insertions(+), 7 deletions(-)
drivers/infiniband/sw/rxe/rxe_loc.h +1 -1
index 0e022ae1b8a5..37484a559d20 100644
@@ -77,7 +77,7 @@ struct rxe_mr *lookup_mr(struct rxe_pd *pd, int access, u32 key,
                         enum rxe_mr_lookup_type type);
 int mr_check_range(struct rxe_mr *mr, u64 iova, size_t length);
 int advance_dma_data(struct rxe_dma_info *dma, unsigned int length);
-int rxe_invalidate_mr(struct rxe_qp *qp, u32 rkey);
+int rxe_invalidate_mr(struct rxe_qp *qp, u32 key);
 int rxe_reg_fast_mr(struct rxe_qp *qp, struct rxe_send_wqe *wqe);
 int rxe_mr_set_page(struct ib_mr *ibmr, u64 addr);
 int rxe_dereg_mr(struct ib_mr *ibmr, struct ib_udata *udata);
drivers/infiniband/sw/rxe/rxe_mr.c +11 -7
index fc3942e04a1f..790cff7077fd 100644
@@ -576,22 +576,27 @@ struct rxe_mr *lookup_mr(struct rxe_pd *pd, int access, u32 key,
        return mr;
 }
 
-int rxe_invalidate_mr(struct rxe_qp *qp, u32 rkey)
+int rxe_invalidate_mr(struct rxe_qp *qp, u32 key)
 {
        struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
        struct rxe_mr *mr;
        int ret;
 
-       mr = rxe_pool_get_index(&rxe->mr_pool, rkey >> 8);
+       mr = rxe_pool_get_index(&rxe->mr_pool, key >> 8);
        if (!mr) {
-               pr_err("%s: No MR for rkey %#x\n", __func__, rkey);
+               pr_err("%s: No MR for key %#x\n", __func__, key);
                ret = -EINVAL;
                goto err;
        }
 
-       if (rkey != mr->rkey) {
-               pr_err("%s: rkey (%#x) doesn't match mr->rkey (%#x)\n",
-                       __func__, rkey, mr->rkey);
+       if ((mr->access & IB_ACCESS_REMOTE) && key != mr->rkey) {
+               pr_err("%s: key (%#x) doesn't match mr->rkey (%#x)\n",
+                       __func__, key, mr->rkey);
+               ret = -EINVAL;
+               goto err_drop_ref;
+       } else if ((mr->access & IB_ACCESS_LOCAL_WRITE) && key != mr->lkey) {
+               pr_err("%s: key (%#x) doesn't match mr->lkey (%#x)\n",
+                       __func__, key, mr->lkey);
                ret = -EINVAL;
                goto err_drop_ref;
        }
--

2.25.1

#2 lore raw
In reply to Md Haris Iqbal
Quoted 11 lines · Md Haris Iqbal

In rxe, the access permissions decide which of the lkey/rkey would be set during an MR registration. For an MR with only LOCAL access, only lkey is set and rkey stays 0. For an MR with REMOTE access, both lkey and rkey are set, such that rkey=lkey.

Hence, for MR invalidate, do the comparison for keys according to the MR access. Since the invalidate wr can contain only a single key (ex.invalidate_rkey), it should match MR->rkey if the MR being invalidated has REMOTE access. If the MR has only LOCAL access, then that key should match MR->lkey.

Quoted 7 lines
Fixes 3902b429ca14 ("RDMA/rxe: Implement invalidate MW operations")
---
 drivers/infiniband/sw/rxe/rxe_loc.h |  2 +-
 drivers/infiniband/sw/rxe/rxe_mr.c  | 17 +++++++++++------
 2 files changed, 12 insertions(+), 7 deletions(-)

If the rkey's and lkeys are always the same why do we store them twice in the mr ?

Quoted 41 lines
drivers/infiniband/sw/rxe/rxe_loc.h +1 -1
index 0e022ae1b8a5..37484a559d20 100644
@@ -77,7 +77,7 @@ struct rxe_mr *lookup_mr(struct rxe_pd *pd, int access, u32 key,
 			 enum rxe_mr_lookup_type type);
 int mr_check_range(struct rxe_mr *mr, u64 iova, size_t length);
 int advance_dma_data(struct rxe_dma_info *dma, unsigned int length);
-int rxe_invalidate_mr(struct rxe_qp *qp, u32 rkey);
+int rxe_invalidate_mr(struct rxe_qp *qp, u32 key);
 int rxe_reg_fast_mr(struct rxe_qp *qp, struct rxe_send_wqe *wqe);
 int rxe_mr_set_page(struct ib_mr *ibmr, u64 addr);
 int rxe_dereg_mr(struct ib_mr *ibmr, struct ib_udata *udata);
drivers/infiniband/sw/rxe/rxe_mr.c +4 -6
index fc3942e04a1f..790cff7077fd 100644
@@ -576,22 +576,27 @@ struct rxe_mr *lookup_mr(struct rxe_pd *pd, int access, u32 key,
 	return mr;
 }
 
-int rxe_invalidate_mr(struct rxe_qp *qp, u32 rkey)
+int rxe_invalidate_mr(struct rxe_qp *qp, u32 key)
 {
 	struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
 	struct rxe_mr *mr;
 	int ret;
 
-	mr = rxe_pool_get_index(&rxe->mr_pool, rkey >> 8);
+	mr = rxe_pool_get_index(&rxe->mr_pool, key >> 8);
 	if (!mr) {
-		pr_err("%s: No MR for rkey %#x\n", __func__, rkey);
+		pr_err("%s: No MR for key %#x\n", __func__, key);
 		ret = -EINVAL;
 		goto err;
 	}
 
-	if (rkey != mr->rkey) {
-		pr_err("%s: rkey (%#x) doesn't match mr->rkey (%#x)\n",
-			__func__, rkey, mr->rkey);
+	if ((mr->access & IB_ACCESS_REMOTE) && key != mr->rkey) {

drivers/infiniband/sw/rxe/rxe_mr.c struct rxe_mr *lookup_mr(struct rxe_pd *pd, int access, u32 key, · line 602

Quoted 10 lines
drivers/infiniband/sw/rxe/rxe_mr.c +7 -0
+		pr_err("%s: key (%#x) doesn't match mr->rkey (%#x)\n",
+			__func__, key, mr->rkey);
+		ret = -EINVAL;
+		goto err_drop_ref;
+	} else if ((mr->access & IB_ACCESS_LOCAL_WRITE) && key != mr->lkey) {
+		pr_err("%s: key (%#x) doesn't match mr->lkey (%#x)\n",
+			__func__, key, mr->lkey);
 		ret = -EINVAL;
 		goto err_drop_ref;
 	}

Hurm, I think rxe still has problems here.

By my reading of the spec a FRWR on a l_key should set the variant bits on the l_key as well, but rxe only updates variant bits on rkeys?

I don't understand why it is trying to keep lkey and rkey seperated..

Are you actually using !IB_ACCESS_REMOTE with FRWR? If so how does it behave on mlx5 devices with regard to the variant bits?

Jason

#3 lore raw
Quoted 2 lines

If the rkey's and lkeys are always the same why do we store them twice in the mr ?

They are not always the same currently. If you request remote access they are the same if you don't rkey is set to zero. You could, of course, check both the keys and the access bits but that is not the way it is written currently.

Bob

#4 lore raw
Quoted 5 lines · Robert B
Quoted 2 lines

If the rkey's and lkeys are always the same why do we store them twice in the mr ?

They are not always the same currently. If you request remote access they are the same if you don't rkey is set to zero. You could, of course, check both the keys and the access bits but that is not the way it is written currently.

Storing the rkey instead of checking the flags seems like a weird obfuscation to me..

Jason

#5 lore raw
Quoted 8 lines
Quoted 4 lines
Quoted 1 line

If the rkey's and lkeys are always the same why do we store them twice in the mr ?

They are not always the same currently. If you request remote access they are the same if you don't rkey is set to zero. You could, of course, check both the keys and the access bits but that is not the way it is written currently.

Storing the rkey instead of checking the flags seems like a weird obfuscation to me..

Jason

One always has the choice to always just use the lkey and check the flags. But referring the rkey just uses one memory reference 😊

#6 lore raw
Quoted 11 lines · Robert B
Quoted 9 lines
Quoted 5 lines
Quoted 2 lines

If the rkey's and lkeys are always the same why do we store them twice in the mr ?

They are not always the same currently. If you request remote access they are the same if you don't rkey is set to zero. You could, of course, check both the keys and the access bits but that is not the way it is written currently.

Storing the rkey instead of checking the flags seems like a weird obfuscation to me..

Jason

One always has the choice to always just use the lkey and check the flags. But referring the rkey just uses one memory reference 😊

Have we reached a consensus here as to how to solve this?

This (and the issue created by dual map) has basically caused a regression in RTRS since the 5.15. And we would appreciate it a lot if the fix goes in and is backported.

#7 lore raw
Quoted 12 lines · haris iqbal
Quoted 10 lines · Robert B
Quoted 9 lines
Quoted 5 lines
Quoted 2 lines

If the rkey's and lkeys are always the same why do we store them twice in the mr ?

They are not always the same currently. If you request remote access they are the same if you don't rkey is set to zero. You could, of course, check both the keys and the access bits but that is not the way it is written currently.

Storing the rkey instead of checking the flags seems like a weird obfuscation to me..

Jason

Quoted 7 lines
Quoted 1 line

One always has the choice to always just use the lkey and check the flags. But referring the rkey just uses one memory reference 😊

Have we reached a consensus here as to how to solve this?

This (and the issue created by dual map) has basically caused a regression in RTRS since the 5.15. And we would appreciate it a lot if the fix goes in and is backported.

I think your patch is close, it should just be tweaked a bit.

Jason

#8 lore raw
Quoted 16 lines · Jason Gunthorpe
Quoted 14 lines · haris iqbal
Quoted 11 lines · Robert B
Quoted 9 lines
Quoted 5 lines
Quoted 2 lines

If the rkey's and lkeys are always the same why do we store them twice in the mr ?

They are not always the same currently. If you request remote access they are the same if you don't rkey is set to zero. You could, of course, check both the keys and the access bits but that is not the way it is written currently.

Storing the rkey instead of checking the flags seems like a weird obfuscation to me..

Jason

One always has the choice to always just use the lkey and check the flags. But referring the rkey just uses one memory reference 😊

Quoted 7 lines
Quoted 5 lines

Have we reached a consensus here as to how to solve this?

This (and the issue created by dual map) has basically caused a regression in RTRS since the 5.15. And we would appreciate it a lot if the fix goes in and is backported.

I think your patch is close, it should just be tweaked a bit.

I couldn't conclude from the conversation above what that tweak should be, if a conclusion has been reached. If not, then I'll wait.

Quoted 2 lines

Jason

#9 lore raw
Quoted 19 lines · haris iqbal
Quoted 18 lines · Jason Gunthorpe
Quoted 16 lines · haris iqbal
Quoted 11 lines · Robert B
Quoted 9 lines
Quoted 5 lines
Quoted 2 lines

If the rkey's and lkeys are always the same why do we store them twice in the mr ?

They are not always the same currently. If you request remote access they are the same if you don't rkey is set to zero. You could, of course, check both the keys and the access bits but that is not the way it is written currently.

Storing the rkey instead of checking the flags seems like a weird obfuscation to me..

Jason

One always has the choice to always just use the lkey and check the flags. But referring the rkey just uses one memory reference 😊

Have we reached a consensus here as to how to solve this?

Quoted 8 lines
Quoted 5 lines
Quoted 3 lines

This (and the issue created by dual map) has basically caused a regression in RTRS since the 5.15. And we would appreciate it a lot if the fix goes in and is backported.

I think your patch is close, it should just be tweaked a bit.

I couldn't conclude from the conversation above what that tweak should be, if a conclusion has been reached. If not, then I'll wait.

The 'rkey' input can be an lkey or rkey, and in rxe the lkey or rkey have the same value, including the variant bits.

So, don't check based on the flags, if the rkey is not 0 check it, otherwise check the lkey

Since we already did a lookup on the non-varient bits to get this far, the check's only purpose is to confirm that the wqe has the correct variant bits.

Jason

#10 lore raw
Quoted 33 lines · Jason Gunthorpe
Quoted 19 lines · haris iqbal
Quoted 18 lines · Jason Gunthorpe
Quoted 16 lines · haris iqbal
Quoted 11 lines · Robert B
Quoted 9 lines
Quoted 5 lines
Quoted 2 lines

If the rkey's and lkeys are always the same why do we store them twice in the mr ?

They are not always the same currently. If you request remote access they are the same if you don't rkey is set to zero. You could, of course, check both the keys and the access bits but that is not the way it is written currently.

Storing the rkey instead of checking the flags seems like a weird obfuscation to me..

Jason

One always has the choice to always just use the lkey and check the flags. But referring the rkey just uses one memory reference 😊

Have we reached a consensus here as to how to solve this?

Quoted 8 lines
Quoted 5 lines
Quoted 3 lines

This (and the issue created by dual map) has basically caused a regression in RTRS since the 5.15. And we would appreciate it a lot if the fix goes in and is backported.

I think your patch is close, it should just be tweaked a bit.

I couldn't conclude from the conversation above what that tweak should be, if a conclusion has been reached. If not, then I'll wait.

The 'rkey' input can be an lkey or rkey, and in rxe the lkey or rkey have the same value, including the variant bits.

Quoted 6 lines

So, don't check based on the flags, if the rkey is not 0 check it, otherwise check the lkey

Since we already did a lookup on the non-varient bits to get this far, the check's only purpose is to confirm that the wqe has the correct variant bits.

Thanks Jason. I sent the patch with the changes suggested. Also, I stole your comment and used it as commit message, since that described the changes accurately. :)

Quoted 2 lines

Jason